Fabric Textile Defect Detection, By Selection A Suitable Subset Of Wavelet Coefficients, Through Genetic Algorithm

Narges Heidari, Reza Azmi & Boshra Pishgoo
International Journal of Image Processing (IJIP), Volume (5) : Issue (1) : 2011 25
Fabric Textile Defect Detection, By Selecting A Suitable Subset
Of Wavelet Coefficients, Through Genetic Algorithm
Narges Heidari narges_hi@yahoo.com
Department Of Computer
Islamic Azad University, ilam branch
Ilam, Iran
Reza Azmi azmi@alzahra.ac.ir
Alzahra University
Tehran, Iran
Boshra Pishgoo boshra.pishgoo@yahoo.com
Alzahra University
Tehran, Iran
Abstract
This paper presents a novel approach for defect detection of fabric textile. For this purpose, First,
all wavelet coefficients were extracted from an perfect fabric. But an optimal subset of These
coefficients can delete main fabric of image and indicate defects of fabric textile. So we used
Genetic Algorithm for finding a suitable subset. The evaluation function in GA was Shannon
entropy. Finally, it was shown that we can gain better results for defect detection, by using two
separable sets of wavelet coefficients for horizontal and vertical defects. This approach, not only
increases accuracy of fabric defect detection, but also, decreases computation time.
Keywords: Fabric Textile Defect Detection, Genetic Algorithm, Wavelet Coefficients.
1. INTRODUCTION
In loom industry, it's very important to control quality of textile generation. Though, loom
machines improvement, decreases defects in fabric textile, but this process can't perform quite
perfect [1]. In the past, human resource performed defect detection process in factories but this
tedious operation, depended on their attention and experiment. So we couldn't expect high
accuracy from them. In other words, according to experimental results that were shown in Ref [1],
human eyesight system can recognize just 50-70% of all defects in fabric textile. In the other
hand, Ref [2] claimed that human's accuracy in fabric defect detection isn't higher than 80%. So,
it seems that automatic evaluation of fabric defects is very necessary.
Because of improvement in image processing and pattern recognition, automatic fabric defect
detection can presents an accurate, quick and reliable evaluation of textile productions. since
fabric evaluation is very important in industrial applications, Different researches were done on
fabric textile, surface of wood, tile, aluminum and etc [2,3,4,5,6]. [3] presented a method for fabric
defect detection based on distance difference. In this method, user can adjust some parameters,
according to the type of textile. [7] evaluated effects of using regularity and local directions as two
features for fabric textile defect detection. [8] used Gabor filters for feature extraction and
Gaussian Mixture model for detection and classification of fabric defects. [9] used structure
characters of fabrics for defect detection and [10] was done this process by using adaptive local
binary patterns.

Many defect detection systems that were presented up now, work on offline mode. In this mode, if
occur some defects in a part of the fabric textile because of machine's problem, all parts of the
fabric will be defective. But in online mode, since defect detection operation perform on textile
generation time, we can stop generation process as soon as occurs a defection. The goal of this
paper is to presents a method for online processing with acceptable accuracy and speed in defect
detection.
Here, we used wavelet filter for fabric defect detection. This filter puts special information in each
quarter of filtered image. In two dimensional wavelet transform, initial image divides to four parts.
The first part is an approximation of initial image and the other parts include components with
high frequencies or details of image, so vertical, horizontal and diagonal defects can be found in
part 2, 3 and 4 respectively. There is a problem for using wavelet transform in fabric defect
detection. The wavelet coefficients that are suitable for a special fabric, may aren't suitable for the
other one. So in this paper, we used separable wavelet coefficients sets for each fabric.
After extracting all wavelet coefficients from a perfect fabric, we found a suitable subset of these
coefficients using genetic algorithm. Then applied this suitable subset to other images and gain
the available defects in each of them. In fact, our final goal is to design a system that can
recognize defective and perfect fabrics. This paper is organized as follows: in Section 2 and 3, we
briefly introduce Wavelet transform and Genetic algorithm. Section 4 describes the proposed
method for fabric defect detection and Our experimental results is shown in Section 5.
2. WAVELET TRANSFORM
Wavelet transform analyses the signal by a group of orthogonal functions in the form of Ψm,n(x),
that computed from translation and dilation of the mother wavelet Ψ. These orthogonal functions
are shown in Equation 1.
(1)
m and n are real numbers in this equation. Analysis and synthesis formulas are shown in
Equation 2 and 3 respectively.
(2)
(3)
Coefficients of this transform are obtained recursively. A synthesis of level j are shown in
Equation 4.
(4)
In two dimensional wavelet transform, initial image in each level, divides to four parts. The first
part is an approximation of initial image and the other parts include components with high
frequencies of image and show edge points in vertical, horizontal and diagonal directions.
Figure1-(a) and (b) show this partitioning for the first and the second order of DWT on a 256 X
256 pixels image. After applying the first order of DWT on image, it is divided to four images IHH,
ILL, ILH and IHL where H and L are high and low bounds respectively [11,12]. This operation can be
performed For each of these four images repeatedly that the result is shown in part (b). Figure 2
shows a real example and the result of applying the first order of DWT on it in part (a) and (b)
respectively.

ILL ILH
ILLLL ILLLH ILHLL ILHLH
ILLHL ILLHH ILHHL ILHHH
IHL IHH
IHLLL IHLLH IHHLL IHHLH
IHLHL IHLHH IHHHL IHHHH
(a) (b)
FIGURE 1: decomposition of an 256 x 256 pixels Image for the (a) first order, (b) second order of DWT
(a) (b)
FIGURE 2: (a): an image, (b): the result of applying the first order of DWT on (a)
Wavelet transform evaluates the difference of information in two different resolutions and creates
a multi-resolution view of image's signal [11], so these transformation are useful for fabric defect
detection.
3. GENETIC ALGORITHM
Genetic Algorithm is a stochastic optimization method. One of the main characteristic of this
algorithm is that it leaves a set of best solutions in population. This algorithm has some operators
for selecting the best solutions, crossing over and mutation in each generation [13]. GA starts
with an initial and random population of solutions. Then a function that called fitness function, are
calculated for each solution. Among initial population, solutions that their fitness values are higher
than the others, are selected for next generation. In the other hand, Crossover and mutation
operators add new solutions to next generation. this operation performs repeatedly to at the final,
the algorithm converges to a good solution. In this paper, we used genetic algorithm for finding a
suitable subset of wavelet coefficients of a special fabric textile.
The simplest form of GA involves three types of operators: [14]
• Selection: This operator selects chromosomes in the population for reproduction. The
fitter the chromosome, the more times it is likely to be selected to reproduce.
256
128

• Crossover: This operator exchanges subsequences of two chromosomes to create two
offspring. For example, the strings 10000100 and 11111111 could be crossed over after
the third locus in each to produce the two offspring 10011111 and 11100100. This
operator roughly mimics biological recombination between two single-chromosome
(haploid) organisms.
• Mutation: This operator randomly flips some bits in a chromosome. For example, the
string 00000100 might be mutated in its second position to yield 01000100. Mutation can
occur at each bit position in a string with some probability, usually very small (e.g.,
0.001).
the simple GA works as follows: [14]
1. Start with a randomly generated population of N L-bit chromosomes (candidate solutions
to a problem).
2. Calculate the fitness F(x) of each chromosome x in the population.
3. Repeat the following steps (a)-(c) until N offspring have been created:
a. Select a pair of parent chromosomes from the current population, with the
probability of selection being an increasing function of fitness. Selection is done
with replacement," meaning that the same chromosome can be selected more
than once to become a parent.
b. With probability Pc (the crossover probability), cross over the pair at a randomly
chosen point (chosen with uniform probability) to form two offspring. If no
crossover takes place, form two offspring that are exact copies of their respective
parents.
c. Mutate the two offspring at each locus with probability Pm (the mutation
probability), and place the resulting chromosomes in the new population.
4. Replace the current population with the new population.
5. If the stopping condition has not satisfied, go to step 2.
4. FABRIC DEFECT DETECTION
Our proposed method has two phases. the First phase includes extracting all wavelet coefficients
from a perfect fabric image and the second one includes finding a suitable subset of all
coefficients using genetic algorithm. This subset can delete main fabric of image and indicate
defects of fabric textile. So after these two phases, we can apply the suitable subset to the other
images and gain the available defects in each of them.
Jasper in [15] showed that if the optimization problem in Eq. 5 can be solved, we can find a
suitable subset of wavelet coefficients.
Where
(5)
In this equation, J function is the normal form of high frequency components of image, n is the
number of coefficients, c is a subset of wavelet coefficients as a solution and p is an array of
image's pixel values. Our goal is to find a c subset that minimize J function. For this purpose, we
can use genetic algorithm. In this algorithm, each solution c is a chromosome and different fitness
functions can be used. In this paper we evaluated some fitness functions but sum of squares and
entropy functions had the best results.
Using a wavelet coefficients subset as a filter that filters images in both horizontal and vertical
directions, has not good results for real fabrics that aren't symmetric. So by using two separable

filters for horizontal and vertical directions, we can obtain the better results. If W1 and W2 be
horizontal and vertical filters respectively and U be a perfect image, then the filtered image F, can
be shown as Equation 6.
(6)
Now we must calculate a fitness function for filtered image F. as mentioned previously, in this
paper, we used sum of squares and entropy functions. These functions can be shown in Eq. 7
and 8 respectively.
(7)
(8)
In Eq. 7, n and m are the number of lines and columns in image and F(i , j) is the pixel value of
image, in line i and column j. in Eq. 8, n is the number of gray levels that here is 256 and d(i) is
probability of gray level i in filtered image F. this probability is a real number between 0 and 1. For
example if 1024 pixels among all pixels have the same gray level value like 100, then can be
written: d(100) = 1/64.
With these fitness functions in Eq. 7 and 8, the optimization problem that were shown in Eq. 5,
can be written in form of Eq. 9 and 10 respectively. As mentioned previously, our purpose is to
minimize J function.
(9)
(10)
The entropy function optimizes wavelet coefficients based on monotony of background. So if the
entropy function value be small, the image will be more monotonic. But the sum of squares
function optimizes image by minimizing its pixel values. furthermore entropy function has simpler
computation in comparison with sum of squares. So we continued our experiments with entropy
function as fitness function of GA.
5. EXPERIMENTAL RESULTS
In this paper, for evaluating the proposed technique, we used two image databases that are
called Tilda [16] and Brodatz [17]. Tilda includes 8 groups of different fabric textile types and
Each group has 50 images with 512 x 768 pixels. Brodatz includes 112 types of fabric images
with 640 x 640 pixels.
Figure 3 shows a defective fabric textile of Tilda database and the filtered image in part (a) and
(b) respectively. The wavelet coefficients of each quarter was optimized by GA while the fitness
function of GA was entropy. This Figure shows that the entropy function has good quality,
because the background is monotonic and defects are appeared clearly.

(a) (b)
FIGURE 3 : (a) a defective fabric textile, ( b) filtered image by suitable coefficients for each quarter
In this experiment, first a perfect image from each fabric used for training phase and a suitable
subset of its wavelet coefficients for each quarter was obtained. Figure 4 shows the histogram of
quarter 4 of Figure 3-(b). this Figure shows a normal distribution. Most of pixels that are related to
background are placed in the middle of distribution but defects are shown in the end of each
littoral. So by selecting two thresholds in proximity of these two littoral, we can perform a
thresholding process on the image and obtain a clear image for place detection of defects. The
image can be more clear by applying some denoising and normalization processes on it.
FIGURE 4: histogram of quarter 4 of Figure 3-(b)
Figure 5-(a) is a fabric textile with diagonal defects. Part (b) of this Figure shows quarter 4 of the
filtered image that diagonal defects was appeared in it. Part (c) and (d) show the result of
applying thresholding and denoising processes on part (b) and (c) respectively. Now part (d) of
Figure 5 has suitable quality for defect detection.

(a) (b)
(c) (d)
FIGURE 5: (a) a fabric with diagonal defects, (b) the quarter 4 of filtered image, (c) the result of applying
thresholding process on part (b), (d) the result of applying denoising process on part (c)
Figure 6-(a) is a fabric textile with horizontal defects. Part (b) of this Figure shows quarter 3 of the
filtered image that horizontal defects was appeared in it. Part (c) shows the result of applying
thresholding on part (b), but this result is not suitable for denoising because the size of noise and
defect areas are same nearly. So in this example we applied sliding window method on image in
part (b) of Figure 6. This method builds 2 windows with size and for each pixel of
image, Then calculates the standard deviation of each window and assignes the result of
Equation (11) to central pixel as new value of it. In this Equation T is a suitable threshold.
New value of Central pixel = (11)
Part (d) of Figure 6 shows this result. In this image, size of defect areas are very larger than size
of noise areas. So denoising process can be done on part (d). the result of denoising process is
shown in part (e). Now part (e) of Figure 6 has suitable quality for defect detection.

(a) (b)
(c) (d)
(e)
FIGURE 6: (a) a fabric with horizontal defects, (b) the quarter 3 of filtered image, (c) the result of applying
thresholding process on part (b), (d) the result of applying sliding window method on part (b), (e) the result
of applying denoising process on part (d)
In this experiment, first a perfect image from each fabric used for training phase and a suitable
subset of its wavelet coefficients for each quarter, was obtained. Then in the test phase, the
other perfect or defective images evaluated by our proposed technique. Table 1 and 2 show
these results for Tilda and Brodatz databases, respectively. Since genetic algorithm starts with a

random initial population, solutions that are produced by each population, must be same or close
to each other. This characteristic is called repetition's capability. We tested this capability and its
results were shown in Table3. In this table, for each quarter of two different fabric types, mean
and standard deviation of ten runs with different initial populations were shown.
Fabric
Type
Image
number
Defective
image
number
Perfect
image
number
Detected
defective
image
number
Detected
perfect
image
number
Accuracy
of correct
detection
C1r1 18 12 6 11 6 94.4%
C1r2 22 14 8 14 7 95.4%
C2r3 20 16 4 15 4 95%
TABLE 1 : Fabric defect detection results for Tilda database
Fabric
Type
Image
number
Defective
image
number
Perfect
image
number
Detected
defective
image
number
Detected
perfect
image
number
Accuracy
of correct
detection
D77 100 23 77 19 76 95%
D79 100 27 73 23 70 93%
D105 100 56 44 51 39 90%
TABLE 2 : Fabric defect detection results for Brodatz database
Fabric type Quarter number Mean Standard deviation
C1r1e3 2 5.9208 0.1092
C1r1e3 3 6.0158 0.1286
C1r1e3 4 5.9466 0.1297
C2r2e2 2 5.7401 0.1607
C2r2e2 3 5.8144 0.1707
C2r2e2 4 5.7530 0.1356
TABLE 3 : results of repetition's capability
6. CONCLUSION
As mentioned previously, in this paper we used a novel method for fabric textile defect detection.
This process had two phases. the First phase includes extracting all wavelet coefficients from a
perfect fabric image and the second one includes finding a suitable subset of all coefficients using
genetic algorithm. The experimental results showed that our method have good accuracy for
defect detection.

REFERENCES
1. K. Srinivasan and P.H. Dastor and P. Radhakrishnaihan, and S. Jayaraman “FDAS: A
Knowledge-based Frame Detection Work for Analysis of Defects in Woven Textile
Structures”, Journal of Textile Institute, vol. 83, pp. 431-447, (1992).
2. R. Chin, “Automated Visual Inspection Techniques and Applications: A Bibliography”,
Pattern Recognition, 15(4): pp. 343–357, (1982).
3. Z. Guanng and W. Jianxia, “ Fabric Defect Detection Method Based on Image Distance
Difference”, Electronic Measurement and Instruments, (2007), pp. 822 -825.
4. D. Chetverikov and A. Hanbury, “Finding Defects in Texture Using Regularity and Local
Orientation “ , Pattern Recognition. 35(10), pp. 2165–2180, (2002).
5. X.Z. Yang and G. Pang and N. Yung, “ Discriminative Fabric Defect Detection Using
Adaptive Wavelets”, Optical Engineering, pp. 3116-3126, December (2002).
6. X.Z. Yang and G. Pang and N. Yung, “ Robust Fabric Dfect Detection and Classification
Using Multiple Adaptive Wavelets”, Image and Signal Processing, IEE Proceedings, volume
152, PP. 712-723 , (2005)
7. D. Chetverikov “ Measuring the Degree of Texture Regularity”, In Proceedings of
International Conference on Pattern Recognition, 1984, volume 1, PP. 80–82, (1984).
8. Yu.zhang,zhaoyang Lu,Jing Li,"Fabric Defect Detection & classification using Gabor filters &
Gaussian Mixture Model",Springer-LNCS,pp:635-644,(2010)
9. HAN leil,LI zhong,"quick defect detection based on structure character of woven fabric
image & wavelet transform",computer engineering & design",(2009)
10. Rong Fu,Meihong Shi,Hongli Wei,Huijuan chen,"fabric defect detection based on adaptive
local binary patterns",international conference on robotics & bi omimetics,pp:1336-
1340,(2009)
11. S. Mallat,“ A Wavelet Tour of Signal Processing”, Academic Press, 2nd ed., San Diego,
(1999)
12. S. Mallat,“A Theory for Multiresolution Signal Decomposition: the Wavelet Representation”,
IEEE Transactions on Pattern Anal. and Mach. Intell., vol. 11, no. 7, pp. 674-693, (1989)
13. Z. Michalewicz, “Genetic Algorithms + Data Structures = Evolution Programs”, AI Series.
Springer-Verlag, New York, 3rd edition, (1996)
14. M. Mitchell, "Genetic Algorithm: An Overview", pp: 31-39, (1995).
15. L.W. Leung and B. King an, “Comparison of Image Data Fusion Techniques Using Entropy
and INI” . 22nd Asian Conference on Remote Sensing, November (2001)
16. Workgroup on"Texture Analysis of DFG, “TILDA Textile Texture Database” ,
http://lmb.informatik.uni-freiburg.de/ research/dfg-texture/tilde

17. P.Brodatz, Textures: A Photographic Album for Artists and Designers , Dover, New York,
(1966)

Fabric Textile Defect Detection, By Selection A Suitable Subset Of Wavelet Coefficients, Through Genetic Algorithm

More Related Content

What's hot (18)

Viewers also liked (14)

Similar to Fabric Textile Defect Detection, By Selection A Suitable Subset Of Wavelet Coefficients, Through Genetic Algorithm (20)

Recently uploaded (20)

Fabric Textile Defect Detection, By Selection A Suitable Subset Of Wavelet Coefficients, Through Genetic Algorithm